Unitree Italy distribution chain

Cobots AMR Quadrupeds Humanoids Accessories Full catalog Price list
Prices Find the right robot Financing
Manufacturing & Logistics Universities & Research
About us Book a call

Research · arXiv · Updated 9 Oct 2026

Unitree in open research: 12 arXiv papers to read before your PoC

Abra Robotics, a selection for labs and R&D
In short: this is not a peer-reviewed review. It is a practical list of preprints we use to align the expectations of customers and universities on the Go2/G1 (and, by extension, on the AS2 / H2). Each entry shows what the paper does, the arXiv link and why it matters for buyers. We re-checked titles, authors and figures against the arXiv abstracts on 9 October 2026.

On arXiv, preprints that use Unitree robots are numerous and growing every month. Here we filter only papers that (a) use real Unitree hardware or a faithful sim, (b) have a takeaway for those buying or integrating in the United Kingdom and Ireland, (c) can be cited in a project document.

For labs and universities: Universities and Research. Ranges cited: Go2, AS2, G1, H2. Product guides: AS2, H2, ROS 2.

Quadrupeds (Go2 → implications for AS2)

Locomotion and sim-to-real

  1. AcL: Fault-Tolerant Quadruped Locomotion (2503.21401): RL teacher-student; limping gaits with up to 4 joints failed in training, validated on a real Go2 with one and two joints failed. Why it matters: it defines robustness acceptance tests for inspection.

  2. MoE + RoboGauge (2602.00678): Mixture-of-Experts and predictive sim-to-real metrics; Go2 on snow, sand, stairs and 30 cm obstacles, up to 4 m/s in speed tests. Why it matters: a framework for choosing policies before risky physical trials.

  3. RANDPOL (2505.19054): Policy with few trainable parameters on Go2, transferred zero-shot to the real robot. Why it matters: useful if your lab has limited GPUs but wants end-to-end RL.

  4. Bridge the Gap (2510.13488): Go2 trained in MuJoCo on a 13.24 m oscillating bridge and tested on the real bridge; more stable with vertical perturbations. Why it matters: walkways and vibrating floors in plants.

Datasets and navigation

  1. Kine2Go (2606.14433): 800 Go2 kinematic trajectories from 40 policies. Why it matters: it accelerates imitation learning without collecting everything from scratch.

  2. Library navigation Go2 EDU (2606.03340): ROS 2 + RTAB-Map + Nav2 in a real library: 100% success in a static scene, 88% with many people, mean map error 3.7 cm. Why it matters: an indoor blueprint for the AS2 EDU/Pro.

  3. ForEnt (2606.19675): Dataset on Go2 in eight forests in the United Kingdom: around 1.7 km covered and 69 cases of the robot getting stuck in vegetation. Why it matters: if your use case is vegetated outdoor, it changes the SKU and safety case.

  4. VOP-Nav crowded navigation (2607.15036): Crowded navigation with LiDAR and velocity obstacles, tested on a real Go2 indoors and outdoors. Why it matters: warehouses and trade shows with people.

Humanoids (G1 → implications for H2 / H2 Plus)

  1. Power model G1 arm (2606.15915): Electrical power model of the 7-degrees-of-freedom left arm: R² 0.933 in identification, 0.965 at unseen speeds. Why it matters: energy-aware planning on long manipulation.

  2. Whole-body teleop G1 (2605.12347): IMU motion capture suit that drives a G1 in real time, validated in MuJoCo and then on the robot. Why it matters: a latency/safety checklist for HRI demos on the H2.

  3. Foundation policy fine-tuning G1 (2512.01358): Fine-tuning with contact and depth signals: on the G1 the "Pick Apple to Bowl" task goes from 48% to 94% success. Why it matters: a realistic context for H2 Plus / GR00T.

  4. What Matters in Humanoid Motion Tracking (2607.19903): Empirical study with the open-source YAHMP framework on G1. Why it matters: it avoids overfitting to a single training setup.

Recommended extra: Action Agent (2605.01477) for language-guided navigation (64.7% of tasks completed on a real G1); the NVIDIA Isaac GR00T documentation, the platform the H2 Plus is based on.

How to use this list

  • Before choosing: decide which paper you want to reproduce. The robot the authors used (Go2 EDU, G1 EDU) indicates the minimum configuration.
  • In the tender: cite 2–3 references as a technical basis, not as a performance guarantee.
  • In the PoC: use the paper metrics (success rate, power R², falls) as indicators to measure on your setup.

Frequently asked questions

Do the paper results also apply to the AS2 and H2?

Not directly. Almost all the papers use the Go2 and G1: the methods and metrics can be reused, but the numbers must be re-measured on your robot.

Do you need an EDU version to reproduce these papers?

Usually yes. The authors use the Go2 EDU or G1 EDU because secondary development is only available on the EDU versions.

Are arXiv preprints reliable?

They are papers not yet peer-reviewed, or not in their final version. Check the figures and experimental conditions in the PDF before using them in a project.

Disclaimer

Preprints are not final peer-reviewed papers; figures and claims should be checked on the PDF and on your setup. Abra Robotics supplies the Unitree range in the United Kingdom and Ireland, with support in English: we sell hardware and integration services; we are not the authors of the cited papers.

Want to set up a lab that reproduces one of these papers? Contact us or start the assessment.

Sources

  1. Xu et al., AcL: Action Learner for Fault-Tolerant Quadruped Locomotion Control, arXiv:2503.21401: arxiv.org/abs/2503.21401
  2. Wu et al., Toward Reliable Sim-to-Real Predictability for MoE-based Robust Quadrupedal Locomotion, arXiv:2602.00678: arxiv.org/abs/2602.00678
  3. Liu, Jain, Nguyen, RANDPOL: Parameter-Efficient End-to-End Quadruped Locomotion via Randomized Policy Learning, arXiv:2505.19054: arxiv.org/abs/2505.19054
  4. Stasica et al., Bridge the Gap: Enhancing Quadruped Locomotion with Vertical Ground Perturbations, arXiv:2510.13488: arxiv.org/abs/2510.13488
  5. Pałucki et al., Kine2Go: Kinematic dataset for the Unitree Go2 robot with diverse gaits and motions, arXiv:2606.14433: arxiv.org/abs/2606.14433
  6. Li et al., Autonomous Navigation System for Library Service Robot Based on Unitree Go2 Edu, arXiv:2606.03340: arxiv.org/abs/2606.03340
  7. Kirdwichai, Tarapore, ForEnt: A Multi-Modal Dataset for Characterizing Quadruped Robot Entrapments in Forest Environments, arXiv:2606.19675: arxiv.org/abs/2606.19675
  8. Wu et al., Learning Agile Navigation in Crowded Environments for Quadruped Robots, arXiv:2607.15036: arxiv.org/abs/2607.15036
  9. Deniz et al., Identification of a Physics-Based Electrical Power Consumption Model for the Unitree G1 Humanoid Arm, arXiv:2606.15915: arxiv.org/abs/2606.15915
  10. Durrani, Khan, Real-Time Whole-Body Teleoperation of a Humanoid Robot Using IMU-Based Motion Capture with Sim2Sim and Sim2Real Validation, arXiv:2605.12347: arxiv.org/abs/2605.12347
  11. Park, Kee, Oh, Modality-Augmented Fine-Tuning of Foundation Robot Policies for Cross-Embodiment Manipulation on GR1 and G1, arXiv:2512.01358: arxiv.org/abs/2512.01358
  12. Amadio, Mingo Hoffman, What Matters in Humanoid General Motion Tracking? An Empirical Study, arXiv:2607.19903: arxiv.org/abs/2607.19903
  13. Sam et al., Action Agent: Agentic Video Generation Meets Flow-Constrained Diffusion, arXiv:2605.01477: arxiv.org/abs/2605.01477
  14. NVIDIA, Isaac GR00T: developer.nvidia.com/isaac/gr00t